Learning Morphology with Pair Hidden Markov Models
Alexander Clark · 2001
In this paper I present a novel Machine Learning technique based on Pair Hidden Markov Models, a statistical model used in bioinformatics. This technique can be used to learn finite-state string to string transductions. I present a model of the acquisition of the English past tense. The same model can also learn without modification the Arabic broken plural, a much more complex morphological system. I also show how this model can be used for unsupervised learning of morphology, and in fact can learn morphology from sets of words automatically induced from unlabelled corpora. I then discuss various other applications and extensions of this technique. 1. Introduction In section 2 I introduce various learning problems in morphology and discuss previous work. In section 3 I introduce Pair Hidden Markov Models. I give an informal discussion of their origins. I then present a formal definition together with various examples. In section 4 I present various algorithms that operate on ...